This was one of the biggest turning points in the build. I stopped treating every finding as if it belonged in the same bucket. That is lazy, and it leads teams to trust weak signals far more than they should.
In A11Y Cat the finding model separates deterministic failures, corroborated failures, state-limited review, visual composition review, human judgement review, and advisory notes. That is not taxonomy for the sake of taxonomy. It changes how the output is read. A confirmed failure is not the same thing as a prompt for human review, and the product should never pretend otherwise.
The technical guide goes much deeper here. It maps how findings carry WCAG references, evidence, limitation reasons, and confidence boundaries. That is the difference between a result dump and a tool that is trying to be responsible with the truth.
I care about this because accessibility work gets damaged by fake certainty. If the tool cannot explain why it thinks something is wrong, or how far that claim really goes, then the tool is adding noise.
Documentation trail
README.mdTECHNICAL_GUIDE.md
